Some things I don't understand with MCP vs. CLI

Permissions: you can already scope CLI command permissions

Discoverability: you can already discover commands/arguments with a help flag

Why add a whole other process to the computer?

Do MCP proponents imagine in the future every program on your PC has another MCP process also running?

It seems like adding extra complexity/moving parts for not much benefit.

It doesn't have to run locally; it can also run on a remote server. I'm about to start using Claude Code + MCP to try and free up my hands.

The MCP can be hosted on the server side, e.g. Atlassian does this now, and with Dynamic Client Registration you can OAuth to the MCP nicely. This is the way forward for MCPs; you can use oauth2-proxy and nginx to e.g. wrap an open source MCP with a layer of SSO and present it to your company if there's a need to host one that isn't run on the provider's infra.

It’s less about the protocol and more about the state. If an agent fails at step 50, a CLI won't help you rollback the context. We’re building Matrix Origin to bring Git-like branching and snapshots to agent memory—making these "100 processes" actually auditable and recoverable.

i don't think so i don't use mcp anymore